https://web.archive.org/web/20251125224147/https://blog.trai...

Archived copy here: https://archive.is/tIXt7

It's more a viewpoint thing. Any construct cryptographers find that runs in constant time is something that could be optimized to run faster for non-cryptographic code. Constant-time constructs essentially are optimizer bug reports. There is always the danger that by popularizing a technique you are drawing the attention of a compiler contributor who wants to speed up a benchmark of that same construct in non-cryptographic code. So maybe it's not intended as sabotage, but it can sure feel that way when everything you do is explicitly targeted to be changed after you do it.

It’s not intentional. The motivations of CPU designers, compiler writers, and optimizers are at odds with those of cryptographers. The former want to use every trick possible to squeeze out additional performance in the most common cases, while the latter absolutely require indistinguishable performance across all possibilities.

CPUs love to do branch prediction to have computation already performed in the case where it guesses the branch correctly, but cryptographic code needs equal performance no matter the input.

When a programmer asks for some register or memory location to be zeroed, they generally just want to be able to use a zero in some later operation and so it doesn’t really matter that a previous value was really overwritten. When a cryptographer does, they generally are trying to make it impossible to read the previous value. And they want to be able to have some guarantee that it wasn’t implicitly copied somewhere else in the interim.

“Sabotage” can be used in a figurative sense that doesn’t insinuate intent. An adjacent example is “self-sabotage”, which doesn’t imply intent.

I advocated for that, and then I completely lost track of the status.

The whole design is ridiculous.

As a concrete example, say I have a (very naive and bad) password-checker that works like this pseudocode:

  > for i = 1 to len(real_password) {
  >   if entered_password[i] != real_password[i] {
  >      return FAILURE
  >   }
  > }
  >
  > return SUCCESS 

Taking this general idea and expanding it, there are lots of places where the timing of branches of code can leak information about some secret, so in cryptographic code in particular, it’s often beneficial to be able to ensure that two branches (the success and failure branches in the above) take exactly the same amount of time so the timing doesn’t leak information. So to fix the above you would probably want to do two things. Firstly set a boolean to failure and still continue the checking to ensure the “return failure quickly” problem doesn’t leak information and also change your password check to check against a fixed-width hash or something so the length of the password itself wasn’t a factor.

The problem is lots of performance optimizations (pipelining, branch prediction etc) work specifically against this goal- they aim to take branches quickly in the happy path of the code because normally that’s what you want to ensure optimal performance.

So say instead of the above I do

  > bool status = SUCCESS
  > for i = 1 to hash_length {
  >   if hash_of_entered_password[i] != hash_of_real_password[i] {
  >      status = FAILURE
  >   }
  > }
  >
  > return status 

In general people want the ability to tell the compiler that they want a particular piece of code to run in constant time. At the moment, in the general case I think you have to break into inline assembly to achieve this.

It turns off CPU features that could cause execution time to vary in a way that depends on the data being operated on.

Or a pragma? Like how OpenMP did it?

AFAIU, cmov wasn't originally intended to be a guaranteed constant-time operation, Intel and AMD won't commit to keeping it constant-time in the future, but it just so happened that at one point it was implemented in constant-time across CPUs, cryptographers picked up on this and began using it, and now Intel and AMD tacitly recognize this dependency. See, e.g., https://www.intel.com/content/www/us/en/developer/articles/t...

> The CMOVcc instruction runs in time independent of its arguments in all current x86 architecture processors. This includes variants that load from memory. The load is performed before the condition is tested. Future versions of the architecture may introduce new addressing modes that do not exhibit this property.

I mean the possibility that the rest of the program guarantees that the address is valid if the condition is true but otherwise it might be valid or invalid. This is probably not important for most applications, but I don't know if there are some unusual ones where it would matter.